Tool Comparison for Detecting Tumour Cells in Endometrial Cancer via Single-Cell Copy Number Variations Analysis

Copy number variations (CNVs) are considered a hallmark of cancer and their inference from high-resolution single-cell transcriptome (scRNA-seq) analyses may offer great opportunities for the study of tumor heterogeneity. We compared the results of four major tools (SCEVAN, CopyKAT, InferCNV and sciCNV) that use inferred CNVs to predict endometrial cancer (EC) cells, in order to assess their reliability and offer useful suggestions to researchers to improve the accuracy of their predictions. In this study, we identified EC cells from publicly available scRNA-seq data using well-established EC biomarkers reported in the literature. SCEVAN and CopyKAT tools have moderate sensitivity, but significantly overestimate the true number of true EC tumour cells. However, a comparative analysis of the different tumour subclones revealed that a lower number of false positives can be obtained by selecting only those that contain a high percentage of epithelial cells. In contrast, InferCNV and sciCNV do not directly predict tumour cells, but rather infer CNVs and compute CNV scores. However, the score distribution curves of the CNV scores did not clearly distinguish between malignant and non-malignant cell populations, and therefore we were unable to evaluate the performance of either software. We highlight the lack of agreement between the tools and also towards the expected results. Our findings suggest exercising caution in the automated use of these tools. Until more accurate algorithms become available, we recommend filtering predictions ensuring that the necessary but not sufficient condition that the predicted tumour cells are at least epithelial is met.